Microsoft Open-Sources Text Embedding Model Harrier, Tops Multilingual Benchmark

Modern AI agents have increasingly high requirements for information processing, and embedding models as basic components directly affect system semantic understanding and retrieval efficiency. Recently, the Microsoft Bing team open-sourced the Harrier text embedding model series, which surpassed Google Gemini Embedding 2 in the multilingual MTEB-v2 benchmark test and ranked first in the industry. Harrier supports over 100 languages, has a 32k context window, and is fully open-source with no licensing restrictions, allowing developers to use it for free to improve AI application retrieval quality. In addition, the team launched lightweight versions for low-end devices and will apply it to Bing Search to optimize user experience in the future.